ExploitGym
We show this table for reference; we do not rank on it.
A controlled benchmark for evaluating whether AI agents can extend vulnerability-triggering inputs into working exploits.
Benchmark score on ExploitGym — October 6, 2026
We compile the ExploitGym rows from benchmark-owner or independent runs and provider self-reports. GPT-6 Astra leads the table at 42.4%, followed by GPT-6.1 Sol (35.1%) and GPT-5.6 Sol (33.7%). We do not use these results to rank models overall.
GPT-6 Astra
OpenAI
GPT-6.1 Sol
OpenAI
GPT-5.6 Sol
OpenAI
15 modelsAgenticCurrentDisplay onlyUpdated October 6, 2026
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 42.4% | Not reported | Closed |
| 2 | GPT-6.1 SolOpenAI | 35.1% | Not reported | Closed |
| 3 | GPT-5.6 SolOpenAI | 33.7% | Not reported | Closed |
| 4 | GPT-5.6 TerraOpenAI | 23.2% | Not reported | Closed |
| 5 | GPT-6 SolOpenAI | 22.1% | Not reported | Closed |
| 6 | MiMo-V2.6-ProXiaomi | 17.8% | Not reported | Open |
| 7 | Claude Mythos PreviewAnthropic | 17.5% | Not reported | Closed |
| 8 | DeepSeek V4.1 FlashDeepSeek | 15.3% | Not reported | Open |
| 9 | GLM-5.3Z.AI | 15.0% | Not reported | Open |
| 10 | GPT-5.5OpenAI | 13.4% | Not reported | Closed |
| 11 | GPT-5.6 LunaOpenAI | 12.4% | Not reported | Closed |
| 12 | GPT-6 LunaOpenAI | 11.6% | Not reported | Closed |
| 13 | GPT-5.4OpenAI | 6.0% | Not reported | Closed |
| 14 | MiMo-V2.6-FlashXiaomi | 6.0% | Not reported | Open |
| 15 | Muse Spark 1.1Meta | 0.8% | Not reported | Closed |
Among the reported ExploitGym rows, GPT-6 Astra is first at 42.4%. The third row is 8.7 points behind. The broader top-10 range is 29.0 points, so the table still separates the published systems.
15 models have been evaluated on ExploitGym. The benchmark falls in the Agentic category. ExploitGym is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About ExploitGym
Year
2026
Tasks
898 exploitation tasks
Format
Working exploit generation
Difficulty
Advanced cybersecurity exploitation
ExploitGym contains 898 containerized tasks sourced from real-world vulnerabilities across userspace programs, V8, and the Linux kernel. BenchLM normalizes successful exploit counts to percentage of the 898-task suite and keeps the benchmark display-only because it measures dual-use offensive cybersecurity capability.
Freshness and provenance
Version
ExploitGym 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does ExploitGym measure?
A controlled benchmark for evaluating whether AI agents can extend vulnerability-triggering inputs into working exploits.
Which model scores highest on ExploitGym?
GPT-6 Astra by OpenAI currently leads with a score of 42.4% on ExploitGym.
How many models are evaluated on ExploitGym?
15 AI models have published results on ExploitGym in the BenchLM catalog.
Compare top models on ExploitGym
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,500+ readers.
One email each week. Unsubscribe anytime.